Thought about OpenAI's language models o1 and o3 for 5 seconds

  • In September 2024, the company presented a preview of the o1 language model.
  • OpenAI announced the o3 and o3-mini models.
    • The previous model is called o1.
    • The company deliberately skipped one digit “out of respect for Telefonica (a telecom company operating under the O2 brand) and the tradition of coming up with bad names,” said OpenAI CEO Sam Altman.
  • Developers trained o3 to “think” before answering a user query.
    • The model responds with a slight delay.
    • It reiterates its reasoning chain before giving the answer it considers most accurate, according to TechCrunch.
  • Users can “adjust” the time for reasoning — the more time given to the model, the better it can handle a request.
  • In the ARC-AGI test, designed to measure an AI’s ability to learn skills beyond its training data, o3 scored 87.5%.
    • A score of 85% is considered “human level.”
    • TechCrunch notes these results should be taken with caution pending external testing.
A comparison of the preview version of o1, o1, and o3 performance:Left chart: a math competitionRight chart: a questionnaire with graduate-level science questionsVertical scale: answer accuracy (source: OpenAI)
  • Currently, both models are unavailable to the general public, but will be open for public safety tests — security researchers will be able to test them.
    • According to Altman, OpenAI plans to release o3-mini by the end of January 2025, followed by o3.
  • OpenAI introduced the preview of the o1 language model (focused on reasoning-intensive tasks) in September 2024.
    • In early December, the company released the full version of o1 and added a ChatGPT Pro subscription with access to an “exclusive” version.